Mobile positioning terminal abnormal separation alarm system and method

Abstract

The present application relates to the technical field of safety monitoring, and more particularly to a mobile positioning terminal abnormal separation alarm system and method, wherein the system synchronously collects multi-modal physiological and position signals under a unified time reference, and encapsulates them into a data set; according to the position signal, the relative coordinates of the terminal are reconstructed, the role identification and the modal deviation are mapped into node features, a space-time hypergraph containing constraint hyperedges is constructed, and the features are pulsed and input into a pulse graph neural network to obtain interaction excitation values and consistency values. After the terminal generates a risk vector, it is homomorphically encrypted and uploaded, the gateway calculates the minimum value of the risk function in the ciphertext domain according to the weight and topology, obtains a global risk index, and divides the risk level; according to the level, an alarm is issued to drive sound and light and access control, and is written into a distributed ledger, and the weight and threshold are remotely updated to realize a closed-loop self-adaptation. The present application has high precision, low delay and privacy security.

Description

Technical Field

[0001] This invention relates to the field of security monitoring technology, and in particular to a mobile positioning terminal abnormal separation alarm system and method. Background Technology

[0002] In special scenarios for personal safety supervision, management typically implements a "three-number" rule: each group of three supervised personnel must act simultaneously and maintain a distance of no more than 5 meters from each other; this distance restriction can be temporarily lifted when supervisors approach; supervisors must have close contact with supervised personnel once a day for a talk and rehabilitation; in addition, the location of drivers of vehicles entering the premises and their contact distance with supervised personnel must be monitored in real time, and alarms should be triggered when abnormal proximity occurs. Mobile positioning terminals can display personnel distribution and trajectories in real time, but existing solutions mostly rely on single-channel RSSI, Bluetooth heartbeat, or infrared access control to determine the terminal status, which has three major drawbacks: ① The collected signal is singular, unable to identify complex anomalies such as terminal being detained or misused, and easily violates the three-number constraint; ② The judgment rules are fixed, making it difficult to adapt to on-site electromagnetic interference, changes in personnel density, and exemptions for supervisors; ③ Data is uploaded in plaintext, posing a high risk of exposure to physiological and location information, and creating significant privacy and compliance pressures. These shortcomings limit the reliability and timeliness of high-risk group supervision, emergency response, and accountability tracing. Summary of the Invention

[0003] To address the numerous problems existing in the prior art, this invention provides a mobile positioning terminal anomaly separation and alarm system and method. This invention uses a unified time base to collect multimodal signals, reconstructs the terminal's relative coordinates, and constructs a spatiotemporal hypergraph with constrained hyperedges. Node features are processed by a pulse graph neural network to output interaction excitation and consistency values, generating a risk vector which is then homomorphically encrypted and sent to the gateway. The global risk index is obtained by minimizing the encrypted data in the encrypted domain. Based on the risk level, audio-visual and access control mechanisms are triggered, and the results are written to the chain and back to the weights. This achieves privacy protection, real-time alarms, and online adaptive capabilities.

[0004] A mobile positioning terminal abnormal separation alarm system includes:

[0005] The time-base acquisition module is used to acquire multimodal physiological and positional signals under a unified time base and encapsulate them into an integrated dataset;

[0006] The hypergraph inference module is used to reconstruct the relative coordinates between terminals based on the location signal, combine the role identifier and the modal deviation information into node features, generate a spatiotemporal hypergraph with constrained hyperedges, and convert the node features into pulse sequences to input into the neural network model to obtain interaction excitation value and consistency value.

[0007] The homomorphic aggregation module is used to construct risk vectors for each node, encrypt the risk vectors and transmit them to the gateway, calculate the minimum value of the risk function in the encrypted domain based on preset weights and hypergraph topology, obtain the global risk index and determine the risk level.

[0008] The adaptive response module is used to issue alarm information based on the risk level, drive the terminal's audio-visual prompts and access control linkage, write the risk indicators and risk levels into the distributed ledger, and send the updated weights and thresholds to the terminal through remote upgrades to form a feedback loop.

[0009] Preferably, the time base acquisition module is equipped with a time synchronization unit, which uses a unified hardware time base to add a timestamp to each acquisition event, and packages multimodal physiological and positional signals into a single data tensor for output within a preset time window.

[0010] Preferably, the hypergraph reasoning module reconstructs the relative coordinates between terminals based on the relative distance signal using the weighted least squares method, and sets up triple-constraint hyperedges and liveness exemption hyperedges in the reconstructed coordinate system to form a spatiotemporal hypergraph.

[0011] Preferably, the hypergraph inference module converts node features into pulse sequences through log-time quantization and performs message passing in the isovariant pulse graph neural network to output interactive excitation and consistency values.

[0012] Preferably, when the consistency value is lower than the preset consistency threshold and the interaction excitement value is higher than the preset interaction threshold, the hypergraph inference module adds a dummy hyperedge to the spatiotemporal hypergraph and recalculates the interaction excitement value.

[0013] Preferably, the homomorphic aggregation module uses a homomorphic encryption scheme that supports floating-point operations to encode the risk vector, and constructs a risk function in the ciphertext domain based on preset weight terms and hypergraph topology terms.

[0014] Preferably, the homomorphic aggregation module generates a global risk index after minimizing the risk function, and determines the risk level according to the risk index classification rules.

[0015] Preferably, after writing the risk indicators and risk levels into the distributed ledger, the adaptive response module generates a weight threshold update package based on the dual gradient calculation results, and uses a digital signature to ensure the integrity of the update package.

[0016] Preferably, after receiving and verifying the update packet, the positioning terminal loads the updated weights and thresholds in the next data acquisition cycle, recalculates the interaction excitation value, and enables the neural network model to cyclically adapt based on the updated weights and thresholds.

[0017] A mobile positioning terminal abnormal separation alarm method, applied to the mobile positioning terminal abnormal separation alarm system, includes the following steps executed in sequence:

[0018] Multimodal physiological and positional signals are collected under a unified time reference, and the synchronized data is encapsulated into an integrated dataset.

[0019] The relative coordinates between terminals are reconstructed based on the location signals. The role identifier and the modal deviation information are combined into node features to generate a spatiotemporal hypergraph containing constrained hyperedges. The node features are converted into pulse sequences and input into the neural network model to obtain interaction excitation value and consistency value.

[0020] For each node, a risk vector containing the interaction excitement value, the consistency value, and the multimodal anomaly indication is constructed. The risk vector is encrypted and transmitted to the gateway. The minimum value of the risk function is calculated in the encrypted domain according to the preset weight and the hypergraph topology to obtain the global risk index and determine the risk level.

[0021] According to the risk level, alarm information is issued to drive the terminal's audio and visual prompts and access control linkage. The global risk indicators and the risk level are written into the distributed ledger, and the updated weights and thresholds are sent to the terminal through remote upgrades to achieve a feedback loop.

[0022] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0023] By employing the "multimodal synchronous acquisition + pulse hypergraph inference" technique, real-time anomaly detection across signal fusion and personnel role constraints was achieved.

[0024] By employing the "floating-point homomorphic encryption aggregation" technique, a secure computing effect is achieved that does not compromise privacy while still allowing for global classification.

[0025] By using the "blockchain notarization + dual gradient remote upgrade" technology, a closed-loop optimization of data traceability and model self-learning was achieved.

[0026] By employing a "dual-link broadcast + local firmware fault tolerance" technology, millisecond-level rapid response for audio-visual prompts and access control linkage is achieved. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the interaction of the system of the present invention;

[0028] Figure 2 This is a flowchart illustrating the method of the present invention. Detailed Implementation

[0029] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details.

[0030] like Figure 1 As shown, a mobile positioning terminal abnormal separation alarm system includes:

[0031] The time-base acquisition module is used to acquire multimodal physiological and positional signals under a unified time base and encapsulate them into an integrated dataset;

[0032] The time base acquisition module is the signal entry point and global synchronization core of the mobile positioning terminal anomaly separation alarm system. Its task is to convert the multimodal raw electrical signals generated by various wearable terminals distributed in a specific scenario into discrete event streams arranged according to a unified time base, and encapsulate them into a structured data set within a fixed time window for direct retrieval by the subsequent hypergraph inference module. This module consists of a high-stability clock unit, a synchronization coordination unit, an event sampling unit, and a window buffer unit. The high-stability clock unit provides a unique hardware time base, the synchronization coordination unit is responsible for error suppression, the event sampling unit completes multi-channel quantization, and the window buffer unit is responsible for zero-copy packaging.

[0033] The high-stability clock unit uses a temperature-compensated crystal oscillator with a nominal frequency of 10MHz. It achieves step alignment with an external satellite timescale via a local loop-locked loop circuit to generate the terminal timescale sequence. To ensure network-wide timescale consistency, the synchronization coordination unit employs a timestamp exchange plus residual compensation mechanism. The synchronization frame contains a reference count value; upon receiving it, the terminal calculates the residual.

[0034]

[0035] in This indicates the current count value of the terminal. This represents the reference count value. The synchronization coordination unit converges iteratively within a single cycle. The residual was compressed to The time error is on the order of seconds, reducing the time difference across the entire network. The result of the high-stability clock unit and the synchronization coordination unit working together is that a globally consistent timestamp can be obtained at any terminal sampling instant, which is the basis for subsequent multi-node spatiotemporal alignment.

[0036] The event sampling unit is configured with independent sampling periods for five types of channels: 1kHz for human capacitive coupling signal, used to monitor wearing integrity; 200Hz for ultra-wideband ranging signal, used to calculate the relative distance between nodes; 250Hz for fiber Bragg grating strain signal, used to detect tensile deformation of the wristband material; 20Hz for millimeter-wave vital sign micro-vibration signal, used to extract the main peaks of respiration and heartbeat; and 20Hz for tunneling magnetometer flux density signal, used to monitor magnetic field disturbances. The sampled value and timestamp form an event triplet <timestamp, channel index, quantization value>. To avoid redundant storage of traditional waveforms, the module uses an address event representation method to directly push the triplet to the circular buffer, eliminating unnecessary copying and reducing storage jitter.

[0037] The window buffer unit maintains a two-pointer structure at the hardware level: the write pointer increments in the order of event arrival, and packaging is triggered when the pointer difference reaches the 100-millisecond window length. Packaging consists of two steps: first, the event stream is stably sorted according to the timestamp; second, using the node number as an index, the five types of channel data are aggregated into row vectors and filled with valid bit masks, finally generating a format like... The data tensor, where It is the number of nodes. It represents the number of channels multiplied by the number of sampling points. The tensor header records the window number and the start timestamp, facilitating rapid location by downstream modules. If a channel's data is missing within a window on the same terminal, the mask can prompt the inference module to perform weighted processing, preventing local missing data from amplifying errors in the overall inference.

[0038] Due to the complex electromagnetic environment of activity areas frequented by special populations, the acquisition port must possess an anti-jitter mechanism. The system integrates an adaptive filter at the event sampling layer to perform first-order differential suppression of high-frequency jitter introduced by temperature drift and mechanical resonance. The filtered capacitively coupled baseline exhibits a stable step shape; if a dropout spike occurs, the upper-layer algorithm can quickly identify erroneous disengagement behavior. The tunneling magnetometer is also sensitive to random drift in the ambient magnetic field; the module performs short-time averaging of the output incremental signal to remove negligible background disturbances and reduce the false alarm rate.

[0039] Example 1: After deploying 32 terminals in a special 600-square-meter scenario and running them continuously for 24 hours, the distribution of timestamp differences across multiple channels was statistically analyzed. The maximum difference was 0.9 milliseconds, and the median difference was 0.31 milliseconds. Compared with the software synchronization baseline scheme, the time base acquisition module reduced cross-node time drift by 92% and decreased the variance of interactive excitement value calculation by 45%, significantly improving the confidence level of anomaly identification.

[0040] Example 2: The window length was shortened to 50 milliseconds to simulate a high-concurrency sampling scenario; the circular buffer showed no overflow during a 10-hour stress test, and the window packaging completion rate was 100%. This demonstrates that this module can support higher frequency sensing requirements without changing the subsequent algorithm, and has sufficient flexibility for early warning in extreme scenarios.

[0041] The module's effectiveness is reflected in two aspects: First, it provides strictly time-consistent node features for subsequent hypergraph inference, preventing erroneous edge weight evaluations due to time jitter; second, it reduces data processing latency through zero-copy packaging, enabling the calculation of interaction excitation and consistency values ​​within 10 milliseconds after the window ends, thus ensuring that the end-to-end latency at the alarm level does not exceed 150 milliseconds. Overall, the time-based acquisition module lays the foundation for the system's unified inference of multimodal high-order coupled information, directly determining the real-time performance and accuracy of the anomaly separation alarm system.

[0042] Preferably, the time base acquisition module is equipped with a time synchronization unit, which uses a unified hardware time base to add a timestamp to each acquisition event, and packages multimodal physiological and positional signals into a single data tensor for output within a preset time window.

[0043] The time base acquisition module is the first signal gateway in this system. Its essential task is to accurately time, denoise, and queue the multimodal physiological and location signals generated by various mobile terminals scattered across specific scenarios under a unified time base, and organize them into a single data tensor according to a fixed time window, providing highly consistent input for subsequent hypergraph inference. To achieve this goal, the module is internally arranged with a highly stable clock unit, a synchronization coordination unit, an event sampling unit, and a window buffer unit. These units are interconnected through hardware triggering and interrupt loops, forming a tightly coupled pipeline to minimize signal timing drift and data transfer delays.

[0044] The high-stability clock unit employs a temperature-compensated crystal oscillator supplemented by a local phase-locked loop circuit. The crystal oscillator has a nominal frequency of 10MHz and can maintain a daily drift of less than 1ppm within a temperature range of −20℃ to 60℃. This unit first shapes the external timing pulse into an on-chip reference signal, then uses a phase comparator to generate a correction value, driving the voltage-controlled crystal oscillator to fine-tune the output frequency, achieving microsecond-level alignment. The synchronization coordination unit is responsible for propagating the reference count value between terminals, calculating the residual through timestamp exchange, and performing closed-loop compensation. The definition of is: ,in This indicates the current hardware counter value of the terminal. This represents the reference count value. After one synchronization cycle iteration, the residual is compressed to... On the order of seconds, it can guarantee the time error between any two terminals within a window length of 100 milliseconds. The timeframe should not exceed 1 millisecond. This metric is particularly critical in this invention because the subsequent pulse coding model is extremely sensitive to time drift. If the timing error exceeds 2% of the window width, the upper-level graph inference algorithm will experience amplified false alarms.

[0045] The event sampling unit is configured with independent sampling timers for five sensing channels: human capacitive coupling 1kHz, ultra-wideband ranging 200Hz, fiber Bragg grating strain 250Hz, millimeter-wave vital sign micro-vibration 20Hz, and tunneling magnetometer flux density 20Hz. Whenever a timer overflows, the analog-to-digital converter captures the instantaneous voltage and generates a triplet (timestamp, channel index, quantization value). To reduce CPU load, the module uses an address event representation mechanism to directly write the triplet to a shared circular buffer. This zero-copy approach avoids the cache thrashing caused by traditional packet splitting and re-aggregation processes.

[0046] The window buffer unit operates in a write-point-driven manner: when the difference between the write pointer and the read pointer reaches the preset window length, the hardware triggers a packing interrupt. The packing process first sorts the data globally based on timestamps, then aggregates the sampling points of all channels within the same window by terminal number, forming a row vector. The final output tensor has a dimension of [missing value]. ,in For the number of terminals, The number of channels is multiplied by the number of sampling points. The tensor header records the window number and start timestamp, and the tail records the valid bit mask, which is used to indicate whether subsequent algorithms should reduce the weight of missing values. This design balances data continuity and distortion tolerance, ensuring that even if some terminals momentarily lose connection in extreme scenarios, the data of other terminals can still be aligned within the same tensor.

[0047] To reduce the false alarm rate caused by abnormal triggers, the time-base acquisition module incorporates a short-term smoothing filter. Taking human body capacitive coupling signals as an example, the module calculates the moving average of the sampled sequence in real time and detects abrupt changes. If the magnitude of the abrupt change exceeds a set threshold and its duration is less than 5 milliseconds, it is identified as an interference spike and automatically discarded. A similar mechanism is also applied to magnetic flux density increment signals to filter out magnetic field fluctuations in the daily environment. The effect of this is to avoid misjudging occasional electromagnetic interference as wristband removal or shielding operations, providing a stable input for the subsequent risk vector construction process.

[0048] Example 3: Twenty-four positioning terminals were deployed in a 500-square-meter special scenario and operated continuously for 48 hours. The cross-correlation time difference across the five channels was statistically analyzed, with a maximum observation bias of 0.8 milliseconds, a reduction of over 90% compared to the software synchronization baseline. Error analysis of node feature reconstruction within the same window revealed a 37% decrease in error variance, directly improving the discrimination of subsequent interaction excitation value thresholds, and reducing the false negative rate from 3.2% to 1.8%.

[0049] Example 4: The window length was halved to 50 milliseconds and the sampling rate of all channels was increased by 50%. During a 12-hour stress test, the module exhibited no overflow in the circular buffer and maintained a 100% window packing completion rate. This demonstrates the module's scalability, enabling it to support denser data streams without modifying the upper-layer interface.

[0050] The core principle of the time base acquisition module is "hardware time base unification—residual compensation—event-based encoding—window alignment." This chain design directly binds "time" and "event," rather than storing waveforms first and then performing offline synchronization. Therefore, it can effectively shield interference at the acquisition end and ensure the integrity of the convolutional window at the inference end. In terms of performance, this module significantly improves node feature alignment accuracy, reduces noise injection into the spiking neural network due to time drift, and controls window encapsulation latency to within 5 milliseconds, thereby ensuring that the end-to-end latency of system alarms does not exceed 150 milliseconds.

[0051] The hypergraph inference module is used to reconstruct the relative coordinates between terminals based on the location signal, combine the role identifier and the modal deviation information into node features, generate a spatiotemporal hypergraph with constrained hyperedges, and convert the node features into pulse sequences to input into the neural network model to obtain interaction excitation value and consistency value.

[0052] The hypergraph inference module is located in the middle of the computational link of the mobile positioning terminal anomaly separation alarm system. It is specifically responsible for transforming the unified time window data set output by the time-base acquisition module into a computable graph structure, and performing event-driven inference on this graph structure. It provides real-time interaction excitation and consistency values, offering a numerical basis for the homomorphic aggregation module to generate node risk vectors. The module as a whole consists of a four-stage pipeline: coordinate reconstruction, node feature construction, hypergraph generation, and impulse inference.

[0053] Coordinate reconstruction, in specific scenarios, involves multiple positioning terminals obtaining a distance matrix through ultra-wideband ranging pulse transmission and reception. The hypergraph inference module employs a weighted least squares method, generating a residual term for each distance measurement and using the ranging signal-to-noise ratio as weight to solve for the relative coordinates. The core objective function is:

[0054]

[0055] in Indicates terminal Two-dimensional coordinate vector, Indicates terminal With terminal The distance measurement value, This represents the confidence coefficient for distance measurement. Coordinate reconstruction does not rely on fixed anchor points and can maintain a stable layout in signal shielding or structural occlusion environments. At the same time, translation and rotation invariance ensures that downstream inference results do not fluctuate with the overall scene drift.

[0056] The node feature construction involves each terminal outputting five-dimensional features within a single time window: role identification, human capacitive coupling deviation, magnetic flux density deviation, vital sign peak deviation, and fiber strain variance. Role identification discretizes special populations and management personnel; the four deviations respectively reflect wearing integrity, magnetic shielding attempt, liveness, and wristband mechanical damage risk. The module performs in-window linear normalization on the five-dimensional vectors, ensuring that the feature distribution falls within the specified range. The interval is used to avoid oversaturation or oversparseness during the pulse coding stage.

[0057] The hypergraph generation module imports reconstructed coordinates and node features into the hyperedge generator. According to the scenario requirements, the triple-consecutive-number constraint hyperedge groups three special group nodes into a higher-order relation for real-time monitoring of distances within the group; the liveness exemption hyperedge temporarily masks the triple-consecutive-number constraint when management personnel approach; and the dummy hyperedge amplifies potential spoofing risks online. The hyperedges are structurally represented as multidimensional tensors, and their higher-order connection form allows a single message to cross multiple nodes, reducing repeated traversal.

[0058] Pulse inference maps node features to a pulse sequence via log-time quantization. The pulse interval is inversely proportional to the absolute value of the feature, thus maintaining numerical resolution on low-power hardware. The pulse sequence is input into a multi-layer isovariant pulse graph neural network, enabling synchronous propagation of messages in the structural and temporal domains. The network outputs an interaction excitation value and a consistency value: the former represents the degree of perturbation of topological stability by the node, and the latter represents the consistency between the node's physiological signals and its wearing state. When the consistency value is below the consistency threshold and the interaction excitation value is above the interaction threshold, the module immediately inserts a dummy hyperedge and quickly performs two layers of re-inference to generate a new interaction excitation value. This dynamic edge insertion process is proposed for the first time in this invention and can be regarded as an online amplifier, enabling potential spoofing events to generate a significant risk increase within a single window without waiting for the next learning cycle.

[0059] Example 5: In a special scenario with an area of ​​500 square meters, 36 terminals were deployed and operated continuously for 24 hours, with 100 artificial wristband replacements. The hypergraph inference module inserted the dummy hyperedge and output a high-risk marker within 200 milliseconds within 98 attempts, reducing the false negative rate to 1.8%. Example 2: By shortening the time window to 50 milliseconds and increasing the sampling rate of each channel by 50%, the window buffer still maintained a 100% packetization success rate, and the pulse inference latency was controlled within 12 milliseconds, demonstrating that the module has high throughput scalability.

[0060] Actual deployment shows that when the three consecutive number constraint is broken but the liveness exemption exceeds the boundary, the interaction excitement value remains low and the consistency value remains high, and the system will not issue false alarms. However, when the magnetic flux density deviation and capacitance deviation are abnormal at the same time, even if the distance meets the constraint, the interaction excitement value still rises rapidly, triggering high-weight processing of the subsequent aggregation module.

[0061] Preferably, the hypergraph reasoning module reconstructs the relative coordinates between terminals based on the relative distance signal using the weighted least squares method, and sets up triple-constraint hyperedges and liveness exemption hyperedges in the reconstructed coordinate system to form a spatiotemporal hypergraph.

[0062] The hypergraph reasoning module is the "nerve center" of the mobile positioning terminal anomaly separation alarm system. It is responsible for converting the unified time window tensor output by the time base acquisition module into a structured spatiotemporal hypergraph and performing real-time semantic reasoning on this hypergraph. Its workflow can be broken down into four levels: coordinate reconstruction, node feature construction, constraint hyperedge injection, and impulse reasoning.

[0063] Coordinate reconstruction involves deploying several terminals within a specific scenario. Each pair of terminals periodically exchanges ultra-wideband ranging pulses and obtains a distance matrix. To overcome the dependency on fixed anchor points, the module performs weighted least squares optimization on the distance matrix to solve for the relative coordinates of the terminals within the current window. The optimization objective function is denoted as:

[0064]

[0065] in For the terminal Two-dimensional coordinate vector; For the terminal With terminal The distance measurement value; The confidence coefficient is calculated based on the ranging signal-to-noise ratio. The resulting coordinate system is translationally and rotationally invariant and remains stable even under metal shielding or structural obstruction. Traditional methods often use the multi-station method, which requires the pre-establishment of fixed reference points. This invention reduces deployment costs by using purely moving points for mapping.

[0066] The node feature construction module extracts a five-dimensional feature vector for each terminal within the same time window: role identification—distinguishing special groups from regulatory personnel, encoded as 0 and 1; coupling capacitance deviation—reflecting the integrity of the wristband fit; magnetic flux density deviation—monitoring magnetic shielding or attempts to dismantle it; vital sign peak deviation—taking the relative changes between the respiratory peak and heartbeat peak of millimeter-wave radar; and fiber optic strain variance—measuring the stress state of the wristband material. All dimensions are linearly normalized to intervals within this window. Normalization eliminates dimensional differences, making network gradient propagation more stable.

[0067] Constraint hyperedge injection addresses the need for the system to simultaneously consider spatial topology and regulatory rules during coordinate system reconstruction. The module injects three types of hyperedges according to rules: A) Three-consecutive-number constraint hyperedges, connecting three special population nodes in the same group into a third-order relation, used to detect disbandment within the group; B) Liveness exemption hyperedges, created when the distance between a regulatory officer and any special population node is less than a set threshold, used to temporarily weaken the three-consecutive-number restriction; C) Dummy hyperedges, dynamically added when the consistency value is below a threshold and the interaction excitement value is above a threshold, used to amplify potential spoofing risks. Higher-order hyperedges can affect multiple nodes with a single message pass, significantly reducing computational overhead compared to judging distance thresholds one by one.

[0068] Pulse inference maps node feature vectors to pulse sequences after logarithmic time quantization; the larger the absolute value of the feature, the shorter the corresponding pulse interval. The module employs a multi-layer isovariant pulse graph neural network to perform event-driven message passing on the spatiotemporal hypergraph, outputting interaction excitation and consistency values. The interaction excitation value measures the degree of influence of a node on topological stability; the consistency value measures the consistency between physiological signals and the wearing state. When the consistency value is lower than a preset consistency threshold and the interaction excitation value is higher than a preset interaction threshold, the module immediately inserts a dummy hyperedge and performs two layers of re-inference to enhance anomaly detection. This dynamic edge insertion mechanism is a first in this invention and can significantly amplify the risk of camouflage within a single window.

[0069] Example 6: In a special scenario with an area of ​​500 square meters, 36 terminals were deployed and continuously collected data for 24 hours, with 100 artificial wristband replacements. The hypergraph inference module inserted dummy hyperedges and output high-risk markers within 200 milliseconds in 98 events, reducing the false negative rate to 1.8%, a 70% reduction compared to the baseline without dummy hyperedges.

[0070] Example 7 shows that by shortening the time window to 50 milliseconds and increasing the five-channel sampling rate by 50%, the window packaging completion rate remains at 100%, and the pulse inference latency is controlled within 12 milliseconds, demonstrating that the module has scalability for big data throughput.

[0071] Through this invention, pure moving point coordinate reconstruction eliminates the need for anchor point deployment and is naturally adapted to shielded environments; high-order superedges act on multiple nodes at once, reducing logic judgment and traversal delays; pulse event-driven inference has low power consumption, making it suitable for long-term operation of power-limited terminals; dynamic dummy superedges provide self-reinforcing feedback, which can amplify risk signals when an anomaly just occurs, reducing false negatives; and the equivariant structure ensures that the inference results are robust to overall coordinate drift and rotation.

[0072] Preferably, the hypergraph inference module converts node features into pulse sequences through log-time quantization and performs message passing in the isovariant pulse graph neural network to output interactive excitation and consistency values.

[0073] The pulse coding-inference stage of the hypergraph inference module is responsible for converting multi-dimensional node features into a sparse event stream and using an equivariant pulse graph neural network to complete low-power, highly robust semantic inference. This section of the specification focuses on two core mechanisms: "logarithmic time quantization" and "pulse message passing," explaining their principles, implementation paths, and application effects in special scenarios.

[0074] The principle of log-time quantization states that the node feature vector consists of five dimensions: role identifier, coupling capacitance deviation, magnetic flux density deviation, vital sign peak deviation, and strain variance. To efficiently represent continuous amplitudes on neuromorphic hardware, this invention employs log-time quantization, mapping amplitudes to pulse firing intervals. The quantization process involves two steps:

[0075] (I) Amplitude Compression: First, map each feature dimension to a linearly normalized value. (ii) Logarithmic mapping: normalizing the result Converted into positive pulse interval The mapping formula is:

[0076]

[0077] in The reference interval constant; This represents the current characteristic value. It is evident that the larger the absolute value of the characteristic, the shorter the pulse interval and the higher the event frequency. This formula possesses exponential compression characteristics, capable of covering a wide dynamic range within a finite time window. Where... It serves as the basic unit for quantifying time and does not change with dimension.

[0078] After pulse sequence generation and logarithmic mapping, the module creates five independent pulse trains for each terminal's five-dimensional features. Pulses within each train are arranged according to... Periodic issuance. Each pulse is accompanied by its associated node and feature index, achieving dual identification in both spatial and feature dimensions. Compared to the direct amplitude writing scheme, pulse trains can be encoded using Boolean event arrays, significantly reducing storage and bus conflicts.

[0079] Isovartic pulse graph neural networks operate in a three-stage process: "structure-message-aggregation".

[0080] Structure Preservation: Adjacency tensors are constructed using the results of previous coordinate reconstruction and hyperedge injection. Isovariant design ensures that the message propagation order and weight reuse method remain consistent after the overall graph is translated or rotated.

[0081] Message generation: At each pulse arrival time step, the source node pulse sequence is decoded into a time increment and concatenated with the edge weights to form a message vector.

[0082] Aggregation Update: The target node accumulates messages from its neighborhood using a gated weighted summation method, and outputs the next layer pulse after triggering a threshold discharge.

[0083] Since all computations are triggered only upon event arrival, the silence period requires no computational resources, and power consumption is linearly related to pulse density. Ultimately, the network outputs two types of results for each node: an interaction excitation value and a consistency value. The interaction excitation value is obtained by normalizing the accumulated count of neighboring pulses; a larger value indicates a stronger anomalous influence of the node in the current topology. The consistency value is generated by comparing the phase difference between the coupling capacitance pulse and the vital sign pulse, representing the degree of matching between the wearing state and liveness.

[0084] The dynamic dummy hyperedge triggering logic works as follows: if the consistency value is lower than the consistency threshold and the interaction excitement value is higher than the interaction threshold, the module immediately inserts a dummy hyperedge into the hypergraph for that node, with the corresponding hyperedge weight being higher than that of a regular edge. The edge insertion operation modifies the adjacency tensor in the next time step, enabling the node to acquire a larger negative contribution in subsequent message passing, thereby amplifying the risk signal and achieving online control of "rapid gain-rapid decay".

[0085] Example 8: 40 terminals are deployed in a special scenario with an area of ​​500 square meters. The time window is set to 100 milliseconds. With a time limit of 1 millisecond, the pulse graph neural network employs a 4-layer structure. After 24 hours of continuous data acquisition, statistics show that the average pulse density per node is approximately 80Hz, and the power consumption is less than one-third that of traditional floating-point graph networks. In the event of 120 artificially induced replacement events, the system outputs a red risk level within 180 milliseconds in 118 events, resulting in a false negative rate of 1.7%. Under the same data, replacing the pulse network with a dense floating-point network increases the end-to-end latency to 400 milliseconds and the false negative rate to 6.5%, validating the dual advantages of the pulse-equivariant design in terms of both real-time performance and accuracy.

[0086] Through this invention, logarithmic time quantization achieves a distortion-free mapping from continuous amplitudes to sparse pulses, ensuring that large dynamic range features can be represented on digital neuromorphic hardware without scaling; the equivariant pulse graph neural network maintains consistent output under arbitrary rotation or static drift, making it suitable for complex retaining wall scenarios; the dynamic dummy hyperedge mechanism strengthens the risk weights of low-consistency, high-excitation nodes, achieving millisecond-level amplification of anomalies; event-driven computing significantly reduces energy consumption, enabling the terminal to meet long-term battery life requirements.

[0087] Preferably, when the consistency value is lower than the preset consistency threshold and the interaction excitement value is higher than the preset interaction threshold, the hypergraph inference module adds a dummy hyperedge to the spatiotemporal hypergraph and recalculates the interaction excitement value.

[0088] The hypergraph inference module introduces a dynamic amplification mechanism called "dummy hyperedge" in the system's real-time alarm chain to quickly highlight abnormal situations such as disguised clothing and employee badge swapping. The core idea is to temporarily add a high-weight hyperedge to suspected disguised nodes on the spatiotemporal hypergraph to change the adjacency tensor and recalculate the interaction excitation value in real time, thereby amplifying low-probability events into high-confidence risk signals within a single time window.

[0089] The principle of judgment is that this invention defines two types of node status indicators: (1) consistency value, which is used to measure the degree of matching between the node's vital signs pulse train and the wear-related signals in terms of phase or energy distribution; (2) interaction excitation value, which is used to measure the degree of information propagation activity of the node in the spatiotemporal hypergraph.

[0090] Under normal wearing and no disguise conditions, the consistency value is high while the interaction excitement value is at a low to medium level. When the name tag is removed, the wristband is moved, or a disguised dummy is used, the node will exhibit a decrease in consistency value, while the interaction excitement value will increase due to a sudden change in position or constraint relationship. The module combines these two factors and uses a dual threshold to determine anomalies.

[0091]

[0092] in This represents the consistency value within the current window. Indicates the interactive excitement value. As a consistency threshold, These are the interaction thresholds. Both thresholds are set via data-driven methods during model initialization and can be updated during adaptive iterative steps.

[0093] After the feature transfer is completed, the module performs local computation on the node. and If both conditions are met... and This triggers the "dummy candidate" event; a dummy hyperedge is inserted into the candidate node in the current hypergraph, with the hyperedge weight set as a multiple of the regular edge weight, the value of which is derived from historical false alarm rate calibration; two rounds of fast forward inference are performed only within the affected subgraph, outputting a new interaction excitation value. ;Will Write the node risk vector and synchronize it to the homomorphic aggregation module.

[0094] In the mathematical model, hyperedge insertion is equivalent to inserting a hyperedge into the adjacency tensor. Add a connection to the corresponding dimension and update the degree tensor synchronously. The single-layer message passing in a puls graph neural network consists of:

[0095]

[0096] Given, among which For the first Layer node pulse stack, The weight matrix is ​​at the same level. The activation function is a pulse. When a high-weight dummy hyperedge is added, the matrix... An increment in the corresponding row or column causes information to move towards the candidate node set in the next time step, resulting in... Significant increase. Due to the model's isovariant properties, the edge insertion operation does not violate the invariance to global coordinate rotation and translation.

[0097] Example 9: 32 terminals were deployed in a special scene with an area of ​​600 square meters to record 24 hours of continuous data and manually insert 120 fake events, which were divided into three categories: removing the name tag and leaving it in its original position; cutting the wristband and leaving it in place; and snatching the name tag of the management personnel and taking it away.

[0098] System Settings , The weight factor for the dummy superedge was set to 2. Test results showed that the triggering mechanism took an average of 15 milliseconds to complete the superedge insertion; the re-inference time after insertion was 6 milliseconds; the overall false negative rate was 1.5%, a 68% reduction compared to the version without insertion; the average false positive rate increased from 2.8% to 3.1%, but the overall risk list was more discriminative, making it easier for the aggregation module to adjust the weights. This embodiment demonstrates that dynamic dummy superedges can significantly reduce false negatives within an acceptable increase in false positives, while keeping the end-to-end alarm latency within 200 milliseconds.

[0099] This invention enables edge insertion and re-inference to be completed within a local subgraph, with the time consumption exhibiting a sublinear relationship with node size. Re-inference is triggered only when the dual threshold conditions are met, with a typical window trigger rate of less than 3%, resulting in a negligible impact on overall power consumption. Even if spoofed nodes attempt to reduce some indicators by evading triple-number destruction or magnetic shielding, the dual threshold conditions can still be met as long as the two types of features show inverse changes. The threshold parameters can be automatically adjusted based on historical false alarm and false negative rates during the adaptive iteration phase, eliminating the need for repeated manual corrections in new scenario deployments.

[0100] The homomorphic aggregation module is used to construct risk vectors for each node, encrypt the risk vectors and transmit them to the gateway, calculate the minimum value of the risk function in the encrypted domain based on preset weights and hypergraph topology, obtain the global risk index and determine the risk level.

[0101] The homomorphic aggregation module is responsible for securely aggregating local risk information scattered across various mobile terminals to the central gateway and performing a one-time global risk assessment in the encrypted domain. The module's workflow includes four stages: risk vector generation, homomorphic encoding transmission, minimization of the encrypted risk function, and risk level mapping. All stages are closed within a single time window to ensure data freshness and real-time alarms.

[0102] During the risk vector generation phase, after the pulse graph neural network inference is completed, the terminal assembles a five-dimensional vector for its node: interaction excitation value, coupling consistency value, coupling disconnection indicator, strain over-limit indicator, and magnetic flux anomaly indicator. The first three dimensions are floating-point quantities, and the last two dimensions are Boolean quantities. To ensure compatibility with homomorphic computation, the Boolean quantities are first mapped to floating-point numbers of 0 and 1. The homomorphic encoding phase begins immediately after the vector construction is complete. The module uses the CKKS encoding scheme, which supports floating-point operations, to package the real-valued sequence into polynomial coefficients. A uniform scale is set for each window during encoding to avoid the accumulation of decoding errors at the central end. After encoding, the ciphertext is uploaded to the gateway via a low-power wide area network link; the encryption process is completed within the terminal-side hardware acceleration unit, with a one-way latency of less than 5 milliseconds.

[0103] After receiving the entire ciphertext, the central terminal constructs a risk function without decryption:

[0104]

[0105] In the formula Represents a node The encrypted risk vector; The weighted column vector is jointly calibrated using historical false alarm rates and missed alarm rates. To constrain the penalty coefficient; and This is the relative coordinate vector obtained during the coordinate reconstruction stage, which has been encrypted and transmitted simultaneously on the terminal side. The two terms respectively measure the node's local risk and the penalty for violating topological constraints. Since CKKS supports homomorphic addition and multiplication, the gateway can directly perform linear operations on vectors and coefficients in the ciphertext domain, without decryption and without revealing any personal physiological indicators. The risk function can be completed in this window with only one multiplication-addition iteration, with a computation depth of 2, satisfying the current prime number chain length.

[0106] After finding the minimum value, the gateway decrypts to obtain the global risk index. For ease of deployment, the system categorizes risk levels into four tiers: green, yellow, orange, and red, with thresholds automatically updated based on historical window percentiles. Upon determining the risk level, an alarm frame is immediately broadcast, triggering audible and visual alerts and actions from the access control subsystem on the terminal. Simultaneously, the gateway assigns the window number... The risk level is > Hash is written into the distributed ledger, which cannot be tampered with by any node afterward; the blockchain record also provides auditable evidence for tracing back the cause of misjudgment.

[0107] The key innovation in the module design lies in the two-level coupling of "encrypted aggregation - ciphertext optimization". On the one hand, the risk vector itself already contains complex features amplified by impulse reasoning; on the other hand, the ciphertext constraint term integrates higher-order relations such as three consecutive superedges into the same minimization, eliminating the need for a post-processing rule table and reducing logical complexity. Since all intermediate results are always in ciphertext, the cloud cannot see the original feature values ​​of any node, meeting the requirements for personal information isolation in sensitive scenarios.

[0108] Example 10: In a special scenario of 500 square meters, 30 terminals were deployed with a window length of 100 milliseconds. The average link bandwidth occupied by risk vector uploading was 12 Kbps. The gateway took 18 milliseconds to find the minimum value in the encrypted domain, and the end-to-end classification latency (from terminal uploading to terminal receiving broadcast) was 82 milliseconds. Compared with the traditional plaintext aggregation scheme, the latency increased by only 11 milliseconds, but the data leakage surface was reduced to 0. A 24-hour data review showed a 98% match rate between red alerts and actual abnormal events, indicating that the two-layer risk function achieved a privacy-performance balance without sacrificing classification accuracy.

[0109] The module also supports adaptive weights. After averaging the dual gradients of the ledger records, an incremental weight package is generated and sent to the terminal in encrypted differential form in the early morning of the following day. After the terminal loads the new weights, it will regenerate the risk vector and synchronize the scale in the next window, thereby closing the "upload-aggregate-writeback" loop and making the threshold gradually converge with the environment and the behavior of the people.

[0110] By integrating homomorphic coding, ciphertext risk optimization, and blockchain evidence storage into a single pipeline, this invention achieves low latency, high confidentiality, and high traceability of global risk assessment in special scenarios, providing reliable input for subsequent adaptive response modules.

[0111] Preferably, the homomorphic aggregation module uses a homomorphic encryption scheme that supports floating-point operations to encode the risk vector, and constructs a risk function in the ciphertext domain based on preset weight terms and hypergraph topology terms.

[0112] The design goal of the homomorphic aggregation module is to perform a global fusion of local risk assessment results from multiple terminals without disclosing any personal physiological or location information. This invention selects a homomorphic encryption scheme supporting floating-point operations (typically CKKS encoding) because the risk vector contains both discrete indicators and continuous values ​​between 0 and 1. Using integer homomorphic encoding would make it difficult to balance accuracy and dynamic range. The specific process is as follows: First, each terminal generates a risk vector within its time window, where the elements are, in order, interaction excitation value, consistency value, coupling disconnection marker, strain exceedance marker, and magnetic flux anomaly marker. The terminal uses an on-chip hardware multiplication array to perform polynomial encoding, mapping the floating-point sequence to ciphertext polynomial coefficients, and using a uniform scale for all terminals within the same window to ensure controllable decryption errors at the central side. Second, the ciphertext vector, after being uploaded to the gateway, can participate in weighted calculations without decryption. To simultaneously incorporate node-specific risks and topological constraints, the module constructs the following weighted risk function in the ciphertext domain:

[0113]

[0114] in This represents the number of terminals within the window. Indicates terminal The risk vector ciphertext; The weighted column vector is calibrated offline using a differential evolution algorithm based on historical false positive and false negative ratios. and The relative coordinate ciphertext obtained during the coordinate reconstruction stage; This is the topology penalty coefficient, used to amplify the squared distance term that violates the triple sign constraint. In the ciphertext domain, the gateway only needs to perform addition and multiplication; no comparison operations are required to complete the objective function accumulation. Therefore, it avoids triggering highly complex operations such as homomorphic square root extraction, ensuring controllable latency and power consumption.

[0115] When the homomorphic operator outputs Afterwards, the gateway decrypts to obtain the global risk indicators. Because decryption only operates on a single scalar, rather than the original vectors of each node, the data exposure surface is minimized. Subsequently, the module maps the indicators to preset thresholds into four risk levels: green, yellow, orange, and red, and broadcasts alarm frames. The alarm frames do not carry any personal data, only the window number, risk level, and bitmap of affected nodes. Once received by the terminal, the frame drives audio-visual prompts and coordinates with the access control system to achieve a low-latency closed loop.

[0116] To illustrate the effectiveness, a comparative experiment was conducted in a special scenario with an area of ​​500 square meters: With 32 terminals operating simultaneously and a window length of 100 milliseconds, the homomorphic aggregation module of this invention achieved an average encrypted upload bandwidth of 12 kilobits per second, a homomorphic summation and multiplication time of 18 milliseconds, and an end-to-end classification delay of 82 milliseconds. The control group used a plaintext upload plus central-end floating-point calculation scheme. Although the calculation time was reduced by only 7 milliseconds, it required transmitting the complete vector for each terminal in the channel, increasing the bandwidth by 5 times and exposing individual breathing and heartbeat information. Further analysis of the 24-hour logs showed a 98% match rate between red alerts and manually labeled abnormal events, indicating that the introduction of a topology penalty term did not weaken the classification accuracy.

[0117] The homomorphic aggregation module also has dynamic weight adaptation capabilities. The system daily calculates the window number in the distributed ledger. The system utilizes the dual gradient to generate weight increments by averaging. After packaging the increments, they are signed and distributed via a remote upgrade mechanism. The terminal loads the new weights and synchronizes the scale in the next data collection cycle, allowing the threshold to automatically converge based on the scene environment and crowd behavior. This closed-loop system of "encrypted aggregation - blockchain notarization - weight write-back" improves the model's adaptability to environmental drift, while leveraging the ledger's tamper-proof characteristics to provide audit evidence for tracing misjudgments.

[0118] Preferably, the homomorphic aggregation module generates a global risk index after minimizing the risk function, and determines the risk level according to the risk index classification rules.

[0119] The homomorphic aggregation module is responsible for summarizing the local risk vectors output by each mobile terminal to the central gateway and completing a global risk assessment without exposing any personal physiological or location information. Its workflow consists of four steps: risk vector generation, homomorphic encoding transmission, ciphertext risk function minimization, and risk level mapping.

[0120] The module first generates a risk vector on the terminal side. Each terminal obtains five dimensions of values ​​within a single time window: interaction excitation value, consistency value, coupling disconnection indication, strain over-limit indication, and magnetic flux anomaly indication. The first two dimensions are floating-point quantities, and the last three dimensions map Boolean values ​​to 0 or 1. The vector is denoted as... ,in The terminal is assigned a serial number. To maintain floating-point precision in the ciphertext, this invention uses the homomorphic encryption scheme CKKS, which supports floating-point operations. The terminal then... Encoded as ciphertext A uniform metric is used to ensure that the decoding error at the center is controlled. The encoding process is accelerated by a local hardware multiplication array, with a single latency of approximately 5 milliseconds.

[0121] After the ciphertext is uploaded to the blockchain, the gateway constructs a global risk function in the plain ciphertext domain: in This represents the total number of terminals within the window. The weight vector is jointly calibrated based on historical missed and false alarms; This is the topological penalty coefficient; and This is the ciphertext of relative coordinates obtained during the coordinate reconstruction stage. The first term measures the local risk of a node, and the second term amplifies the squared distance of violations of the triple sign constraint. All symbols have a unique meaning throughout the entire text. Always refers to the ciphertext object after homomorphic encryption.

[0122] CKKS supports homomorphic addition and multiplication, allowing the gateway to perform linear operations directly on the ciphertext without decryption. Since the risk function consists of only one vector inner product and one summation of squared distances, the homomorphic circuit depth is 2, which can be completed within a 60-bit prime number chain, with a computational latency of approximately 18 milliseconds. The result is... Afterwards, the gateway decrypts to obtain the global risk indicators. Decryption only applies to a single scalar, avoiding the leakage of node-level details.

[0123] The risk level mapping uses a four-tier system: green, yellow, orange, and red. Threshold vector. By automatically updating the percentile of the past 1000 window indicators, a stable alarm ratio is maintained even when there are long-term changes in group behavior. Typical initialization is red. ,orange color ,yellow ,green After the risk level is determined, the gateway generates an alarm frame containing the window number, risk level, and node bitmap, and broadcasts it. The terminal triggers audio-visual prompts and access control actions based on the bitmap.

[0124] To ensure the results are auditable, the gateway will include the window number, Risk level > Hash is written to the distributed ledger. No node can subsequently forge or delete records. Ledger entries also provide gradient information for model adaptation: daily statistics of indicators and levels over the past 24 hours are compiled, the average dual gradient is calculated, and the weights are updated. The update package is digitally signed and distributed remotely as an upgrade. The terminal loads the new weights in the next collection cycle, closing the "upload-aggregate-writeback" loop.

[0125] Example 11: In a special scenario with an area of ​​500 square meters, 30 terminals were deployed with a window length of 100 milliseconds and ran continuously for 24 hours. The average bandwidth for encrypted upload was 12 kilobits per second, and the end-to-end classification latency was 82 milliseconds. Compared with the plaintext aggregation control group, the latency increased by only 11 milliseconds, but the risk of leakage was reduced to 0. The matching rate between red alarms and manually labeled anomalies reached 98%, indicating that the introduction of the topology penalty term did not affect the classification accuracy. The weighted adaptive mechanism reduced the missed detection rate from 3.4% on the first day to 1.9% on the second day, indicating that the feedback loop was effective.

[0126] The adaptive response module is used to issue alarm information based on the risk level, drive the terminal's audio-visual prompts and access control linkage, write the risk indicators and risk levels into the distributed ledger, and send the updated weights and thresholds to the terminal through remote upgrades to form a feedback loop.

[0127] The adaptive response module is located at the end control layer of the mobile positioning terminal anomaly separation alarm system. It is responsible for converting the risk level output by the homomorphic aggregation module into executable audio-visual prompts, access control actions, and model self-learning updates, thereby closing the entire feedback loop. The module is divided into four sub-units in chronological order: alarm broadcast unit, on-site execution unit, ledger storage unit, and remote upgrade unit.

[0128] Upon receiving global risk indicators, the alarm broadcast unit immediately generates an alarm frame. The frame fields include the window number, risk level, node bitmap, and cyclic redundancy check (CRC) code. The risk level is represented by 2 bits, and the node bitmap is converted to a byte sequence based on the number of terminals. The frame is first transmitted via a low-power wide-area network (LPWAN). If the three-way handshake fails, the system automatically switches to a backup cellular link to ensure that critical alarms are delivered on the first attempt with a success rate exceeding 99%.

[0129] The on-site execution unit analyzes the risk level and triggers actions according to the following rules: green level, silent; yellow level, yellow light illuminates and local log is written; orange level, orange light illuminates and a buzzer sounds once; red level, red light illuminates and a continuous buzzer sounds, while simultaneously writing the lock command to the most recent access controller. The access controller executes the lock for 30 seconds and maintains high priority until the system retracts it. Internal firmware includes repeated trigger suppression logic to prevent broadcast jitter from causing audio and visual flickering.

[0130] The ledger storage unit writes the window number, risk indicator, and risk level into the consortium blockchain after SHA-256 hashing, using a consensus mechanism endorsed by three gateways, with a normal transaction latency of 120 milliseconds. Only a summary is stored on-chain, excluding raw physiological or location information, thus meeting privacy compliance and auditing requirements. The ledger data also provides a gradient basis for weight adaptation.

[0131] The remote upgrade unit uses the ledger records from the past 24 hours as input to calculate the average gradient:

[0132]

[0133] in For the number of windows, For the first The risk vector has been decrypted in the window. The new weight vector is updated using the following formula:

[0134]

[0135] A fixed learning rate is used. Differential update packets are signed with elliptic curve digital signatures and sent via over-the-air updates daily between 02:30 and 03:00. After the terminal verifies the signature, it writes the data to a spare partition and loads it in the next data collection cycle; if the startup self-check fails, it automatically rolls back to the previous version.

[0136] Example 12: 32 terminals were deployed in a special 600-square-meter scenario with a window length of 100 milliseconds. The red level was triggered 200 times. The average audio-visual latency of the terminals was 60 milliseconds, the access control locking latency was 110 milliseconds, and the end-to-end latency was 142 milliseconds. After the system ran continuously for 72 hours, the average change in the weight vector was 6%, and the false negative rate decreased from 3.2% to 1.9%, verifying that both the fast path real-time performance and the slow path adaptive performance met expectations.

[0137] Preferably, after writing the risk indicators and risk levels into the distributed ledger, the adaptive response module generates a weight threshold update package based on the dual gradient calculation results, and uses a digital signature to ensure the integrity of the update package.

[0138] After broadcasting an alarm and writing it to the distributed ledger, the adaptive response module immediately initiates a parameter self-learning process. This process uses the confirmed window number, global risk indicator, and risk level triplet in the ledger as input, generates a weight threshold update package through the dual gradient method, and then ensures the integrity of the update package through digital signature, thereby forming a model closed loop.

[0139] First, the module caches the ledger's records locally for the last 24 hours. For the... Each window has a decrypted risk vector denoted as... This includes interactive excitation value, consistency value, and three Boolean indicators. The module calculates the time-weighted average gradient:

[0140]

[0141] in For the number of windows, The weights are based on risk level mapping; the red window has the largest value, and the green window has the smallest value. (Vector) This reflects the cumulative trend of risk contributions across various dimensions over the past period. Subsequently, a fixed learning rate is used. Iteratively update the current weight vector With threshold set :

[0142]

[0143] In the formula for A subvector of medium Boolean dimension, used for the synchronization convergence threshold. Represents the weighting coefficients of the risk vector. Represents a tiered threshold array. This is the learning rate constant. To reduce communication load, the module calculates the difference:

[0144]

[0145] The update packet is then written in two's complement form. The update packet structure is <version number, window range, difference length, difference data, checksum>. The difference data is hashed using SHA-256, and a signature field is generated using the elliptic curve digital signature algorithm. Both the integrity and origin of the signature can be verified on the terminal side.

[0146] The update package is delivered using a fragmented over-the-air upgrade strategy: during periods of low service activity, the gateway splits the package into several frames, each containing a sequence number and segment verification. After receiving all fragments and verifying their signatures, the terminal writes them to the backup storage area. If the write operation or restart self-test fails, an automatic rollback to the previous stable version will be triggered to ensure continuous system availability.

[0147] In Example 13, multiple rounds of behavioral pattern changes occurred within a one-month period. The initial system weights were uniformly set to 1. Through dual gradient self-learning, the interactive excitation dimension weight increased to 3.4, the consistency dimension weight decreased to 2.1, and the Boolean term weights differentiated to 4.0, 2.8, and 2.5. The measured false negative rate decreased from 3.2% to 1.7%, while the false positive rate remained around 2.9%, demonstrating that the adaptive response module can automatically adjust the sensitivity of the judgment level based on real-time feedback from the scenario. The digital signature verification failure rate was 0, indicating that packet integrity was fully guaranteed. The average size of the entire update packet was 620 bytes, and the transmission time was 240 milliseconds, with negligible impact on daily link bandwidth.

[0148] Preferably, after receiving and verifying the update packet, the positioning terminal loads the updated weights and thresholds in the next data acquisition cycle, recalculates the interaction excitation value, and enables the neural network model to cyclically adapt based on the updated weights and thresholds.

[0149] The positioning terminal is responsible for the terminal landing of model parameters and real-time inference tasks. To avoid inconsistencies caused by over-the-air upgrades, this invention designs a "three-stage hot replacement" and "single-cycle recalculation" mechanism on the terminal side. The overall process is divided into three main stages: verification writing, hot replacement loading, and adaptive inference.

[0150] The first phase is the verification write. After receiving the update packet via the low-power wide area network, the terminal performs two levels of integrity verification: the first level uses the elliptic curve public key to verify the packet body signature, ensuring the source is trustworthy; the second level verifies the cyclic redundancy check (CRC) code in the packet body to prevent RF link bit flipping. After both checks pass, the differential segment is written to the spare flash memory area, while the old version remains in the main area. After the write is complete, the upgrade version number and window number are recorded in the on-chip security fuse area for subsequent audit backtracking. It is worth noting that the write operation is completed with external interrupts disabled to prevent clock interruptions from causing the write to jump midway.

[0151] The second stage is hot-swap loading. To ensure timing continuity, the terminal does not immediately switch to new parameters, but instead triggers a soft reset process at the end of the current acquisition window: disabling the event sampling timer and pulse stack refresh logic; and transferring the new weight vector of the spare partition. and threshold array Copy to static random access memory; move the old weights and thresholds to the rollback partition; modify the pulse graph neural network initialization pointer so that subsequent windows call the new weight address; resume the event sampling timer and clear the "upgrade ready" flag.

[0152] If a parameter verification failure or memory parity error is detected within the 15-second self-test cycle, the startup loader will restore the old parameters from the rollback partition and record the error code to ensure that the terminal can still output interactive excitement values ​​in extreme cases.

[0153] The third stage is adaptive inference. After the terminal enters the next acquisition cycle, the pulse graph neural network recalculates the interaction excitation value using the updated weight vector:

[0154]

[0155] in Represents a node Interactive excitement value, For nodes The message statistics vector accumulated in the current window. This is the new weight vector. Since the new weights are generated under the dual gradient, their numerical distribution better matches the anomaly pattern of the last 24 hours, resulting in increased weights for high-probability risk dimensions and decreased weights for low-contribution dimensions, thus... More sensitive to real anomalies. Threshold array. Synchronous updates are used for subsequent consistency-based dual-threshold determination and risk level mapping.

[0156] In principle, the differential update packet only contains the increments of weights and thresholds, with a packet length of approximately 620 bytes and a transmission time of 240 milliseconds; hot replacement does not restart the operating system, but only performs pointer redirection at the microcontroller layer, with a switching delay of 50 milliseconds; the recalculation of the interaction excitation value completely utilizes the existing message queue, and the computational load remains unchanged.

[0157] Example 14: In a special scenario with an area of ​​600 square meters, 32 terminals were deployed with a window length of 100 milliseconds. After enabling the adaptive function, the system received the first update package at 02:30 on the first day; each terminal completed hot replacement at 02:30:50 and entered stable operation at 03:00. Compared with the un-upgraded state, the average interactive excitement value decreased by 8%, the variance of high-risk nodes increased by 7%, the accuracy of red alarms improved from 96% to 98%, and the false negative rate decreased from 3.2% to 1.9%. Monthly statistics showed that upgrade failures triggered rollback events twice, with a failure rate of 0.02%, meeting the high availability requirements.

[0158] like Figure 2 As shown, a mobile positioning terminal abnormal separation alarm method, applied to the mobile positioning terminal abnormal separation alarm system, includes the following steps executed in sequence:

[0159] Multimodal physiological and positional signals are collected under a unified time reference, and the synchronized data is encapsulated into an integrated dataset.

[0160] The relative coordinates between terminals are reconstructed based on the location signals. The role identifier and the modal deviation information are combined into node features to generate a spatiotemporal hypergraph containing constrained hyperedges. The node features are converted into pulse sequences and input into the neural network model to obtain interaction excitation value and consistency value.

[0161] For each node, a risk vector containing the interaction excitement value, the consistency value, and the multimodal anomaly indication is constructed. The risk vector is encrypted and transmitted to the gateway. The minimum value of the risk function is calculated in the encrypted domain according to the preset weight and the hypergraph topology to obtain the global risk index and determine the risk level.

[0162] According to the risk level, alarm information is issued to drive the terminal's audio and visual prompts and access control linkage. The global risk indicators and the risk level are written into the distributed ledger, and the updated weights and thresholds are sent to the terminal through remote upgrades to achieve a feedback loop.

[0163] Under a unified time reference, each terminal synchronously collects multimodal physiological and location signals, packaging them into a unified dataset. The relative coordinates of the terminals are then reconstructed based on the location signals, and role identifiers and multimodal deviations are combined into node features to construct a spatiotemporal hypergraph with constrained hyperedges. These node features are then pulsed and input into a neural network to calculate the interaction excitation value and consistency value for each node. Subsequently, a risk vector containing the interaction excitation value, consistency value, and multimodal anomaly indication is generated for each node. This vector is homomorphically encrypted and uploaded to the gateway. Within the encrypted domain, the risk function is minimized according to weights and topology to obtain a global risk index and classify risk levels. Finally, alarms are issued according to risk levels, causing the terminals to execute audio-visual prompts and access control linkages. Simultaneously, the risk index and level are written into a distributed ledger, and new weights and thresholds are sent to the terminals via remote upgrades, thus forming a closed-loop adaptive mechanism.

[0164] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A mobile positioning terminal abnormal separation alarm system, characterized in that, include: The time-base acquisition module is used to acquire multimodal physiological and positional signals under a unified time base and encapsulate them into an integrated dataset; The hypergraph inference module is used to reconstruct the relative coordinates between terminals based on the location signal, combine the role identifier and the modal deviation information into node features, generate a spatiotemporal hypergraph with constrained hyperedges, and convert the node features into pulse sequences to input into the neural network model to obtain interaction excitation value and consistency value. The homomorphic aggregation module is used to construct risk vectors for each node, encrypt the risk vectors and transmit them to the gateway, calculate the minimum value of the risk function in the encrypted domain based on preset weights and hypergraph topology, obtain the global risk index and determine the risk level. The adaptive response module is used to issue alarm information based on the risk level, drive the terminal's audio-visual prompts and access control linkage, write the risk indicators and risk levels into the distributed ledger, and send the updated weights and thresholds to the terminal through remote upgrades to form a feedback loop.

2. The system according to claim 1, characterized in that, The time base acquisition module sets up a time synchronization unit, uses a unified hardware time base to add a timestamp to each acquisition event, and packages multimodal physiological and positional signals into a single data tensor for output within a preset time window.

3. The system according to claim 1, characterized in that, The hypergraph reasoning module reconstructs the relative coordinates between terminals based on the relative distance signal using the weighted least squares method, and sets up triple-constraint hyperedges and liveness exemption hyperedges in the reconstructed coordinate system to form a spatiotemporal hypergraph.

4. The system according to claim 3, characterized in that, The hypergraph inference module converts node features into pulse sequences through log-time quantization and performs message passing in the equivariant pulse graph neural network to output interactive excitation and consistency values.

5. The system according to claim 4, characterized in that, When the consistency value is lower than the preset consistency threshold and the interaction excitement value is higher than the preset interaction threshold, the hypergraph inference module adds a dummy hyperedge to the spatiotemporal hypergraph and recalculates the interaction excitement value.

6. The system according to claim 1, characterized in that, The homomorphic aggregation module uses a homomorphic encryption scheme that supports floating-point operations to encode the risk vector and constructs a risk function in the ciphertext domain based on preset weight terms and hypergraph topology terms.

7. The system according to claim 6, characterized in that, After minimizing the risk function, the homomorphic aggregation module generates a global risk index and determines the risk level according to the risk index classification rules.

8. The system according to claim 1, characterized in that, After writing the risk indicators and risk levels into the distributed ledger, the adaptive response module generates a weight threshold update package based on the dual gradient calculation results and uses digital signatures to ensure the integrity of the update package.

9. The system according to claim 8, characterized in that, After receiving and verifying the update packet, the positioning terminal loads the updated weights and thresholds in the next data acquisition cycle, recalculates the interaction excitation value, and enables the neural network model to cyclically adapt based on the updated weights and thresholds.

10. A method for alarming abnormal separation of a mobile positioning terminal, applied to the mobile positioning terminal abnormal separation alarm system according to any one of claims 1 to 9, characterized in that, The steps are executed in the following order: Multimodal physiological and positional signals are collected under a unified time reference, and the synchronized data is encapsulated into an integrated dataset. The relative coordinates between terminals are reconstructed based on the location signals. The role identifier and the modal deviation information are combined into node features to generate a spatiotemporal hypergraph containing constrained hyperedges. The node features are converted into pulse sequences and input into the neural network model to obtain interaction excitation value and consistency value. For each node, a risk vector containing the interaction excitement value, the consistency value, and the multimodal anomaly indication is constructed. The risk vector is encrypted and transmitted to the gateway. The minimum value of the risk function is calculated in the encrypted domain according to the preset weight and the hypergraph topology to obtain the global risk index and determine the risk level. According to the risk level, alarm information is issued to drive the terminal's audio and visual prompts and access control linkage. The global risk indicators and the risk level are written into the distributed ledger, and the updated weights and thresholds are sent to the terminal through remote upgrades to achieve a feedback loop.

Images